genmodelfit

Right-size image / video / audio / 3D generation models to your PC. 画像・動画・音楽・効果音・読み上げ・3D の生成モデルが、その PC で動くか・快適かを判定するツールです。

genmodelfit is inspired by llmfit, which does this for LLMs. It is not an official llmfit project or fork. llmfit に着想を得た、生成モデル向けのツールです(llmfit の公式の派生ではありません)。

Status: v0.1.0 (alpha). / 状態: v0.1.0(アルファ版)。

$ genmodelfit recommend --vram 8 --ram 32 -n 2
PC: assumed 8.0 GiB VRAM (--vram), RAM 32.0 GiB (--ram)
Recommended: 12 of 83 (215 hidden: uncensored 24, tools 34, size unknown 4, too tight 153; 25 more
    workflows merged into their models)
== image (画像): top 2 of 43  (-m image for more)
Status      Model                     Use     Prec Score  VRAM Mem% Mode    Disk License    Auto
● Marginal  Z-Image-Turbo             gen+5   int8    54  7.6m  96% GPU~    11.3 apache-2.0 yes
● Marginal  FLUX.2 [Klein] 4B Base    edit    fp8     41   4.5  56% GPU+RAM 11.5 apache-2.0 yes
== video (動画): top 2 of 5  (-m video for more)
Status      Model                     Use     Prec Score  VRAM Mem% Mode    Disk License    Auto
● Marginal  Wan 2.1 1.3B T2V          gen     fp16    50  7.8m  97% GPU~     9.2 apache-2.0 yes
● Marginal  Wan 2.1 ControlNet 1.3B   edit    bf16    39   7.7  96% GPU+RAM 10.6 mit        yes
== music (音楽): top 2 of 9  (-m music for more)
Status      Model                     Use     Prec Score  VRAM Mem% Mode    Disk License    Auto
● Good      ACE-Step 1.5 Turbo        gen+2   ?       65  6.0m  76% GPU~     9.3 mit        yes
● Marginal  ACE-Step 1.5XL Turbo      gen     bf16    40   5.2  65% GPU+RAM 18.5 mit        yes
== sfx (効果音): top 2 of 2  (-m sfx for more)
Status      Model                     Use     Prec Score  VRAM Mem% Mode    Disk License    Auto
● Perfect   stable-audio-3-small-sfx  gen     ?       68   3.6  45% GPU      2.1 other      license
● Good      stable-audio-open-1.0     gen+1   ?       54   4.9  61% GPU~     5.3 other      license
== tts (読み上げ): top 2 of 16  (-m tts for more)
Status      Model                     Use     Prec Score  VRAM Mem% Mode    Disk License    Auto
● Perfect   Chatterbox                gen     ?       80   3.5  44% GPU      2.0 mit        workflow
● Perfect   Qwen3-TTS-12H…CustomVoice gen     ?       80   3.2  40% GPU      1.7 apache-2.0 workflow
== 3d (3D): top 2 of 8  (-m 3d for more)
Status      Model                     Use     Prec Score  VRAM Mem% Mode    Disk License    Auto
● Good      HY 3D 2.0                 gen     fp16    64  6.7m  84% GPU      4.6 other      license
● Good      TripoSplat                gen     fp16    57   5.8  72% GPU      3.5 other      license
Order: runs first, then Auto=yes, then Score (0-100). VRAM: GiB needed, m = measured.
Mode: GPU all on GPU, GPU~ parts swapped in turn, GPU+RAM offload (slow), CPU (slow), - no run.
Mem%: share of VRAM (of RAM for CPU), RAM = system RAM short. Auto: yes, or first failed check.
One row per model: Prec = best precision that fits, +N = other workflows (--all-workflows).
Estimates without m are rough. Details: genmodelfit info <name or id>. Apps: --json (schema.md).
? = unknown (precision, license, size: not judged), - = not applicable (e.g. no run).
Note: 24 uncensored model(s) hidden; add --content all to show them.

Without -m / --task, recommend shows each of the six types (image, video, music, sfx, tts, 3d). One row is one model: workflows that use the same weights (same family and size) in another precision or for another use are folded into it (+5 = five more workflows; --all-workflows lists them), and the best precision that fits is chosen (Prec; like llmfit's best quant). Narrow down inside a type with --task: -m も --task も無ければ 6 つの種類ごとに出します。1 行は 1 つのモデルで、同じ重み(系統と大きさが同じ)を 別の精度・別の使い方で使うワークフローはまとめ(+5 = ほかに 5 つ。--all-workflows で全部)、載る中で一番良い精度を 選びます。種類の中は --task で絞ります。

$ genmodelfit recommend --vram 8 --ram 32 -m image --task edit -n 3
PC: assumed 8.0 GiB VRAM (--vram), RAM 32.0 GiB (--ram)
Recommended: 3 of 18 (303 hidden: other types 162, uncensored 19, tools 19, other uses 63, too
    tight 40; 2 more workflows merged into their models)
Status      Model                     Use     Prec Score  VRAM Mem% Mode    Disk License    Auto
● Marginal  FLUX.2 [Klein] 4B         edit+1  fp8     41   4.5  56% GPU+RAM 11.6 apache-2.0 yes
● Marginal  FLUX.2 [Klein] 4B Base    edit    fp8     41   4.5  56% GPU+RAM 11.5 apache-2.0 yes
● Marginal  Z-Image-Turbo             edit+5  bf16    41   7.5  93% GPU+RAM 22.2 apache-2.0 yes
Order: runs first, then Auto=yes, then Score (0-100). VRAM: GiB needed, m = measured.
Mode: GPU all on GPU, GPU~ parts swapped in turn, GPU+RAM offload (slow), CPU (slow), - no run.
Mem%: share of VRAM (of RAM for CPU), RAM = system RAM short. Auto: yes, or first failed check.
One row per model: Prec = best precision that fits, +N = other workflows (--all-workflows).
Estimates without m are rough. Details: genmodelfit info <name or id>. Apps: --json (schema.md).
? = unknown (precision, license, size: not judged), - = not applicable (e.g. no run).
Note: 19 uncensored model(s) hidden; add --content all to show them.

The default output is a table for people; add --json for apps and agents (llmfit prints JSON by default for recommend; genmodelfit keeps the table so a first try is readable). 既定の出力は人が読む表です。アプリやエージェントは --json を付けてください(llmfit の recommend は既定が JSON ですが、 genmodelfit は初めて試す人のために表のままにしています)。

Why / なぜ作るか

llmfit covers LLMs only (llmfit#41). For diffusion, video and audio models the memory needed changes several times over with precision, resolution, frame count and offloading to system RAM, so they need their own database and scoring.

生成モデルは、精度・解像度・コマ数・メインメモリへの逃がし方で、必要な VRAM が数倍変わります。 そのため LLM とは別のデータと判定が要ります。

Install / 入れ方

Recommended: install it as a command-line tool with pipx or uv. They put genmodelfit on PATH for you (run the PATH command once, then open a new terminal):

pipx install genmodelfit        # once: pipx ensurepath
uv tool install genmodelfit     # once: uv tool update-shell

Or with pip:

pip install genmodelfit

python -m genmodelfit … always works, whatever PATH says (e.g. python -m genmodelfit recommend; on Windows also py -m genmodelfit …). Use it when genmodelfit is "not recognized" / "command not found".

If pip printed "WARNING: The script genmodelfit.exe is installed in '…\Scripts' which is not on PATH", the command was installed but that folder is not on PATH. Common cases:

• Python from python.org installed without "Add python.exe to PATH" → the Scripts folder of that Python is not on PATH.
• pip fell back to a user install (no write access to the system Python) → the command is in %APPDATA%\Python\Python3XX\Scripts (Windows) or ~/.local/bin (Linux), which is often not on PATH.
• Microsoft Store Python → its Scripts folder lives deep under %LOCALAPPDATA%\Packages\… and is not on PATH.
• Installed into a virtual environment that is not activated → activate it (.venv\Scripts\activate / source .venv/bin/activate) or run .venv\Scripts\genmodelfit.

Fix: add the folder named in the warning to PATH (or use pipx / uv above, or python -m genmodelfit). A terminal opened before PATH changed does not see the change: open a new one.

おすすめは pipx か uv でコマンドとして入れる方法です(PATH を通すコマンドを 1 回だけ実行し、端末を開き直す)。 pip で入れた場合も、python -m genmodelfit … は PATH に関係なく必ず動きます(Windows は py -m genmodelfit … も可)。 pip が「The script genmodelfit.exe is installed in '…\Scripts' which is not on PATH」と出したら、コマンドは入っていますが、 その場所に PATH が通っていません。よくあるのは、python.org の Python を「Add python.exe to PATH」なしで入れた、 書き込めないため利用者ごとの場所(Windows は %APPDATA%\Python\Python3XX\Scripts、Linux は ~/.local/bin)に入った、 Microsoft Store の Python、仮想環境を有効にしていない、の 4 つです。警告に出た場所を PATH に足すか、pipx / uv か python -m genmodelfit を使ってください。PATH を変える前に開いた端末には反映されないので、新しい端末を開いてください。

From source / ソースから:

git clone https://github.com/laxshmate/genmodelfit && cd genmodelfit && pip install -e .

Python 3.10+. Windows and Linux. GPUs: NVIDIA (CUDA) in v0.1; other PCs are judged as CPU-only.

Usage / 使い方

Common options / 共通の指定:

• --json — machine-readable output for apps and agents (docs/schema.md) / アプリ向けの JSON
• --vram 12 --ram 32 — judge with other VRAM / RAM sizes; the GPU generation of this PC is kept when it has a usable NVIDIA GPU (docs/hardware.md) / VRAM・RAM の量を変えて判定(この PC に使える NVIDIA の GPU があれば、その世代は引き継ぐ)
• --width 1024 --height 1024, --frames 81, --length 30, --precision fp8 — judge a specific setting / 設定を指定して判定
• -m image,video — types / 種類(カンマ区切り可)
• --task edit — uses inside a type: gen, edit, i2v, lipsync, motion, voice, upscale, tool, or task codes such as t2i (genmodelfit types) / 種類の中の用途
• --all-workflows — one row per workflow instead of one per model / モデルにまとめず、ワークフローごとに出す
• --content sfw|all|uncensored — filter by content mark (below; default sfw) / 内容の印で絞る(既定は sfw)
• --fit perfect|good|marginal|runnable|all — only that level, every level that runs (fit default) or everything / その段階だけ・動くもの全部(fit の既定)・全部
• --min-fit perfect|good|marginal|too_tight — only this level or better (recommend: marginal by default; not together with --fit) / この段階以上だけ
• --sort type|score|vram|disk|name (fit, search) — type (default): type, then fit level, then Score; score: one ranking across types; vram / disk: smallest first / 並べ方
• --include-unknown-license — also show models with a file whose license is unknown (hidden by default) / ライセンスの分からないものも出す
• --license apache-2.0,mit — every file's license must be in the list; unknown never matches / すべてのファイルのライセンスがこの中にあるものだけ
• --comfyui-version 0.40.0 — the ComfyUI version your app runs, for the gate / 使う ComfyUI の版(自動おすすめの判定用)

Order / 並べ方: fit and search show one table, by type (image, video, music, sfx, tts, 3d), then fit level (Perfect → Good → Marginal → ? → Too tight), then Score (0–100: fit level, VRAM headroom, readiness for auto-recommendation, measured or not, popularity; formula in docs/schema.md); auto-recommendable models are not moved up (see the Auto column). recommend shows the top models of each type, with auto-recommendable ones first. The first lines show this PC and how many models were hidden and why (uncensored, tools, license unknown, too tight, …). --help of each command lists preconditions, exit codes and examples. fit・search は 1 つの表で、種類 → 段階 → 点数の順。recommend は種類ごとの上位で、自動でおすすめに出せるものを先に します。先頭に PC と、隠した件数とその理由を出します。

Example / 例 (COLUMNS=100 genmodelfit fit --vram 24 --ram 128, 24 GB GPU / 128 GB RAM assumed; rows cut with …):

PC: assumed 24.0 GiB VRAM (--vram), RAM 128.0 GiB (--ram)
All models: 141 (93 hidden: uncensored 24, tools 34, license unknown 32, too tight 3; 89 more
    workflows merged into their models)
Type  Status      Score Model                    Use     Prec  VRAM Mem% Mode    License    Auto
image ● Perfect      91 SDXL1.0 Base             gen+3   fp16  8.0m  33% GPU     openrail++ license
image ● Perfect      84 FLUX.2 [Klein] 4B Base   edit    fp8   13.2  55% GPU     apache-2.0 yes
image ● Perfect      80 Anima Preview 3 Base     gen     bf16   7.0  29% GPU     other      license
image ● Perfect      80 SDXL Turbo 1.0           gen     fp16   8.2  34% GPU     other      license
image ● Perfect      79 NetaYume Lumina          gen     bf16  11.6  48% GPU     gemma      license
image ● Perfect      79 PixelDiT 1300M           gen     bf16   9.0  38% GPU     other      license
…
video ● Perfect      92 Wan 2.1 1.3B T2V         gen     fp16 13.3m  55% GPU     apache-2.0 yes
video ● Good         67 Wan 2.2 5B TI2V          gen     fp16 15.7m  66% GPU~    apache-2.0 yes
video ● Good         63 Wan 2.1 ControlNet 1.3B  edit    bf16  17.1  71% GPU     apache-2.0 yes
video ● Marginal     44 Wan 2.2 14B I2V          i2v+2   fp8   21.0  88% GPU~    apache-2.0 yes
…
music ● Perfect      94 ACE-Step 1.5 Turbo       gen+2   bf16 10.8m  45% GPU     mit        yes
music ● Perfect      82 ACE Step v1 3.5B         gen+2   bf16   8.7  36% GPU     apache-2.0 yes
…
tts   ● Perfect      81 Qwen3-TTS-12…CustomVoice gen     bf16   5.0  21% GPU     apache-2.0 workflow
tts   ● Perfect      80 Chatterbox               gen     fp32   3.5  14% GPU     mit        workflow
…
3d    ● Perfect      89 HY 3D 2.1                gen     fp16  9.4m  39% GPU     other      license
3d    ● Perfect      88 HY 3D 2.0                gen     fp16  6.7m  28% GPU     other      license
Order: Type, Status, Score (--sort). Score 0-100. VRAM: GiB needed, m = measured.
Mode: GPU all on GPU, GPU~ parts swapped in turn, GPU+RAM offload (slow), CPU (slow), - no run.
Level caps: GPU~ is at most Good; GPU+RAM and CPU are at most Marginal (Mem% can be low).
…

Python:

import genmodelfit as gmf

pc = gmf.system()
for f in gmf.recommend(modality="image", task="edit", limit=5, comfyui_version="0.40.0"):
    print(
        f.model.id, f.model.precision, f.score, f.auto, f.estimate.fit, f.estimate.vram_gb, f.estimate.basis
    )
    print(
        [w["id"] for w in f.workflows]
    )  # other workflows of the same model / 同じモデルのほかのワークフロー
    f.to_dict()  # the same dict as one item of `recommend --json` / `--json` の 1 件と同じ

# the same list as `genmodelfit fit` (runnable only, by type → level → score) / `fit` と同じ一覧
rows = gmf.fit_list(sort="type")  # fit="good", show_all=True, include_tools=True, … like the CLI
gmf.search("flux", license=["apache-2.0", "mit"])

st = gmf.data_status()  # is the model list fresh enough for auto-recommendation? / 一覧の新しさ
if st.expired:  # HEAD checks older than 30 days → auto-recommendations drop / 期限切れ
    print("run: genmodelfit update --from-release", st.expires_at, st.auto_count)

Hardware / ハードウェアの検出

• Only NVIDIA GPUs confirmed by nvidia-smi are used for judging (v0.1). A GPU is judged as CPU when no PyTorch CUDA build supports its generation (Kepler and older) or its driver is too old for all of them (CUDA 12 builds need 525+, CUDA 13 builds need 580+, Blackwell needs 570+). The reason is shown in system and --json.
• On GPUs where ComfyUI computes in fp32 (Maxwell, some Pascal, GTX 16xx, T1000 …) working memory and the VAE are doubled. On GPUs without bf16 (Pascal, Volta, Turing), model families that ComfyUI never runs in fp16 (LTX, Qwen Image, Z-Image, HunyuanVideo, ACE-Step …) also compute in fp32, and so do VAEs without fp16 (SD / Flux KL, LTX).
• The level uses total VRAM; when the VRAM free right now is not enough, a note says to close other apps (vram_free_short).
• When the RAM is short, a fixed-size page file / swap sets how far "slow but runs" goes (system-managed: 1.5× RAM).
• AMD / Intel GPUs are reported (Windows: WMI + registry, Linux: lspci) but never used for judging. Host name, user name, GPU UUID and PCI ids are not collected. Details and sources: docs/hardware.md.

判定に使うのは nvidia-smi で確かめた NVIDIA の GPU だけ。今の PyTorch が動かない世代(Kepler 以前)・古いドライバは CPU として判定し、 理由を出します。AMD・Intel の GPU は知らせるだけです。出典と細かい決まりは docs/hardware.md。

Fit levels / 段階

Like llmfit, the level comes from the utilization of the memory the run mode uses (VRAM on the GPU, RAM on CPU): llmfit と同じく、動かし方で使うメモリの使用率で決めます。

Every estimate says whether it is measured (m after the VRAM figure, a real run) or estimated (formula). Measured values win. Models whose download size is unknown are not judged (shown as ?). すべての見積もりに「実測」か「推定」かを付けます。実測があれば実測を使います。大きさが分からないモデルは判定しません。

Data sources / データ源

The bundled list has 323 entries (image 161, video 114, TTS 20, music 15, 3D 10, SFX 3) that make 204 models. Like llmfit, Hugging Face models are included only with at least 10,000 all-time downloads (1,000 for trending ones), and merges and test stubs are skipped (docs/sources.md). llmfit と同じく、Hugging Face のモデルは全期間のダウンロード数が 10,000 以上のもの(伸びているものは 1,000 以上)だけを入れます。

• ComfyUI workflow templates (MIT): which models a workflow uses, download URLs, hashes, required ComfyUI version, custom nodes. API (cloud) templates are excluded. The index's size field is a rounded, hand-written total, so genmodelfit uses the sum of the file sizes.
• Hugging Face Hub API: licenses (following base_model chains and a table of verified upstreams for repackaged files), gating, file sizes, sha256, and discovery of community models.
• Civitai is not used: its terms limit automated access and forbid reselling compiled data.
• Measurements: src/genmodelfit/data/measurements.json. Contributions welcome (see CONTRIBUTING.md).

Details: docs/sources.md. genmodelfit stores model information only; it never hosts weights. genmodelfit が持つのはモデルの情報だけで、モデル本体は持ちません。

Content marks / 内容の印

• sfw — default.
• uncensored — models whose safety filtering was removed or that are made for adult output (Hugging Face not-for-all-audiences tag, or names such as uncensored / abliterated / nsfw). They are in the list (24 in the bundled list) and marked [U] in tables, but recommend / fit / search and gmf.recommend() / gmf.search() show them only with --content all or --content uncensored (content="all" / "uncensored" in Python). info <id> works for them as usual. The evidence is kept in content_evidence. Whether to show them is up to the app using genmodelfit; they are never auto-recommended by default.
• Models aimed at sexual content involving minors are excluded from the list entirely.

uncensored は、安全フィルタを外したモデルや成人向けの出力を想定したモデルです。一覧には載せ、印を付けて います。年齢制限があるので、recommend / fit / gmf.recommend() の既定(sfw)では出さず、 --content all か --content uncensored を指定したときだけ出します(info <id> はいつもどおり使えます)。表示するかどうかは使う側のアプリが決めます(既定では自動おすすめに出しません)。 未成年の性的表現を目的とするものは一覧から除外します。

Auto-recommendation gate / 自動でおすすめに出す条件

A model is auto: yes only if all machine-checkable conditions hold: every file's license (traced to the upstream, not the uploader's label) is on a commercial-use allowlist (apache-2.0, mit, bsd, cc0, unlicense; replaceable), nothing is gated, an official ComfyUI workflow without custom nodes, the ComfyUI version is supported, every download URL answered an anonymous HEAD within the last 30 days, sha256 is known, the size is known, and the content is sfw. Otherwise it is held with machine-readable reasons (held_reasons in JSON, e.g. license; the table's Auto column shows the first one). In the bundled list 83 of 323 entries are auto. When a license cannot be verified it is treated as unknown and held. See docs/gate.md.

すべての条件を満たすものだけを自動でおすすめに出し、残りは理由付きで保留にします。ライセンスが確かめられない ものは「不明」として保留にします。アプリは使う ComfyUI の版を --comfyui-version で渡してください。

Privacy / 送るもの・送らないもの

• recommend, fit, search, info, system, types, gate and the Python API (gmf.recommend(), gmf.search(), gmf.info(), gmf.system(), gmf.data_status()) make no network access. Hardware is read locally (nvidia-smi, psutil). This is tested with every socket blocked.

• genmodelfit update --from-release only downloads db.json and db.json.sig from this project's GitHub release.

• genmodelfit update only reads public data with GET / HEAD: the ComfyUI workflow templates and releases on GitHub, the Hugging Face Hub API, and the model download URLs (HEAD, plus a ranged GET of the first bytes of the main weight files to read their header — dtype and tensor shapes; no weights are downloaded).

• genmodelfit bench talks only to the ComfyUI URL you give with --url (plus the official template files on GitHub when ComfyUI does not have them).

• Hardware information (GPU name, VRAM, RAM, CPU, disk) is read locally with nvidia-smi and psutil and is never sent. Measurements from bench are saved only on your PC; they are shared only if you open a pull request yourself (CONTRIBUTING.md). Saved measurements hold no host name, user name, paths or GPU UUID.

• No telemetry. No hardware information, measurements, prompts or other user content is ever sent anywhere. Tests replace the PC's details with marker strings and check that no request URL, header or body of update, update --from-release or bench contains them. Requests carry the User-Agent genmodelfit/<version> (+https://github.com/laxshmate/genmodelfit), and never the credentials in your ~/.netrc (proxy and certificate settings such as HTTPS_PROXY are still honored).

• recommend・fit・search・info・system・types・gate と Python の関数はネットにつなぎません (PC の情報は手元で読みます。通信をふさいだ試験で確かめています)。

• update --from-release は、このプロジェクトの GitHub Release から db.json と署名だけを取ります。

• update は公開の情報を読むだけです(GitHub の ComfyUI テンプレート、Hugging Face Hub の API、モデルの URL への HEAD と、 本体のファイルの先頭の見出し(dtype・テンソルの形)だけの Range 要求。モデル本体は落としません)。

• bench は --url で指定した ComfyUI とだけ話します(ComfyUI に無い公式テンプレートは GitHub から取ります)。

• PC の情報(GPU の名前・VRAM・メモリ・CPU・空き)は nvidia-smi と psutil で手元で読むだけで、送りません。 bench の実測は手元に保存するだけで、利用者が自分で Pull Request を出したときだけ共有されます (保存する実測に PC の名前・ユーザー名・パス・GPU の UUID は入りません)。

• 利用状況の送信(テレメトリ)はありません。PC の情報・実測・プロンプトなどを送ることはありません。User-Agent は genmodelfit/<版> で、~/.netrc の資格情報は付けません(プロキシ・証明書の設定は使います)。

Accuracy and limits / 精度と限界

See docs/accuracy.md. In short / 要点:

• Estimates are calibrated on only 9 real runs on one GPU (RTX 5080 16 GB). Leave-one-out error could be evaluated on 4 of them (image and 3D: max 6.4%, mean 4.6%); video and audio could not be evaluated yet. Treat est. values as rough guidance. 実測は 1 種類の GPU で 9 件だけです。1 件ずつ外して当てる方法で評価できたのは 4 件(画像・3D、最大誤差 6.4%、 平均 4.6%)で、動画と音声はまだ評価できていません。
• ComfyUI's dynamic VRAM uses more memory when it is available, so a peak measured on a large GPU is not the minimum requirement.
• Some templates list alternative files (e.g. 1.3B and 14B variants); genmodelfit counts one variant per part, which may not match the variant you pick.
• The HEAD checks in the bundled list expire after 30 days; run genmodelfit update --from-release (or update --heads-only) to refresh them, otherwise auto-recommendations drop to zero. genmodelfit system and gmf.data_status() tell whether the list in use has expired (expired, expires_at). 同梱の一覧の HEAD の確認は 30 日で切れます。切れたかは genmodelfit system と gmf.data_status() で分かります。
• v0.1 counts NVIDIA GPUs only.

License / ライセンス

genmodelfit is licensed under the Mozilla Public License 2.0 (MPL-2.0), and so is the bundled data (db.json, measurements, calibration). MPL-2.0 is a file-level copyleft: if you distribute modified genmodelfit files, you must make those files' source available under MPL-2.0; you may use genmodelfit unmodified (or import it) in a larger work under any license, including closed-source apps. When you distribute genmodelfit inside your app (for example in a PyInstaller exe), keep its license notices and tell your users where its source code is available (https://github.com/laxshmate/genmodelfit, MPL-2.0 §3.2); the same applies to certifi (MPL-2.0). アプリに同梱して配るときは、ライセンスの表示を残し、genmodelfit(と certifi)のソースの入手先を利用者に知らせてください。 Parts of the data come from ComfyUI workflow templates (MIT) and the Hugging Face Hub API; see THIRD_PARTY_NOTICES. Model licenses are reported, not granted.

genmodelfit と同梱のデータは MPL-2.0 です。genmodelfit のファイルを直して配るときは、そのファイルのソースを MPL-2.0 で公開してください。直さずに使う(import する)だけなら、ソース非公開のアプリにも組み込めます (ライセンスの表示は残してください)。データの出典の表示は THIRD_PARTY_NOTICES にあります。 モデルのライセンスは表示するだけで、利用の許諾はしません。各モデルのライセンスを確かめてください。

This Source Code Form is subject to the terms of the Mozilla Public
License, v. 2.0. If a copy of the MPL was not distributed with this
file, You can obtain one at https://mozilla.org/MPL/2.0/.